Trucks with fiber optic cables, have taken root in my neighborhood, like terrible pecking birds pulling worms.

I knew trouble was coming, when the engender showed up, and marked the neighborhood up with orange paint, on snow.

The first time, it snowed, and covered up the markings, the second time, the snow melted, like my hope of connectivity.

I cracked AI today, I had AI, create wonderful summaries of my poems.

I didn't check what it had to say about my special poems, like, "Run Child, Run From The Bus, The School Is Not For Us".

This a major breakthrough for me, as I have spent about 10 weeks, studying local AI, and ensuring proper models and luxury software.

To run the famous gpt-oss-20b I just say: gpt-oss blanced-moe, and the llama.cpp loads the model with the correct configuration.

I also have qwen coder with the latest release of llama.cpp, and a bunch of other things, audio to text, text to image, all local.

odor, as in catpea odor (based on muriel), is my latest catpea blog generator, and you guessed it; it supports AI - what a time to be alive.

I say odor-agent standard-catpea-profile description, and it connects to local AI, sends it my poem.

An system prompt, that states: "You are a careful editor."

And an agent question: "Write a one-paragraphs summary of this post."

My local AI performed beautifully, summarizing poems in a beautiful way.

It is pretty clear that those 20GB thinking files, as I call them, actually hold a perfectly smart AI.

This is just about where I finally understood, how to make AI even smarter.

How to lift it from the strange world of Large Language Models, into the more human world of programming.

Let me teach you how to escape non determinism, and chaotic AI behavior and make AI learn.

You know how I send my poem to my local AI, to have it summarized?

To make AI learn, really, really learn, you add another step after that.

"You say, take a look a what you have done, and tell me if perhaps you can do it better."

So we make AI self evaluate its response, fixing a million bugs, again these really are thinking machines.

And then we take what it said about, how it can do it better,

And we add that to, "Write a one-paragraphs summary of this post."

So we asking AI to self evaluate, and create a memory of good ideas.

This is why some cheap AI may seem clunky today, they are not yet very smart, they haven't learned their lessons yet.

I took this to another level, I am currently testing three programming languages AI can actually understand.

And my next experiment will be asking AI, to not just improve the question...

But this thinking harness of self evaluation, what else can it add, to improve its thinking.

Planning in steps, by creating a first step, evaluating the data the program produces.

And creating a next step, so this is not simple planning, this is planning with program execution.

The next thing it will invent after that, is just abandoning code and sticking it self in the flow as a command.

What a terrible day, for my internet company, to be splicing fiber optic cable.

And wouldn't you know it, it is already 9:20 PM I am sure they left without turning the inter back on.